Software is spending money.
October 2026
Agents now buy data, rent compute, book services and pay other agents. Payment rails assume the agent is trusted. Expense tools assume the spender is human. Nothing governs what machines spend, and nothing proves it afterward.
The situation
The volume is arriving before the rules.
Visa, Mastercard, Stripe, Google, OpenAI and Shopify have all shipped ways for software to pay without a human in the loop. One purchase can flow through a cart protocol, a signed authorization, an agent check and a card settlement.
The protocols don't compete, they stack, and more will arrive while others are shelved. Each one fails in its own dialect, and none keeps a record that anyone else can verify.
Allow, deny, or ask.
We believe every payment should get a verdict before it moves. Authority over money should be specific, limited and revocable, set by the people who answer for the risk, and changeable without touching the agent's code. Every payment should leave a receipt that is written when it happens, can't be edited, and can be checked by a stranger.
A payment that cannot be evidenced should be treated as one that didn't happen.
What we're building
One gate for agent money. Rules before, proof after.
We never hold funds
We don't hold funds, and we never cross a reversibility boundary without permission.
Open by default
Our client library is open source under Apache 2.0, works without us, and keeps your payments and receipts on your machine.
One vocabulary
We're publishing one open vocabulary for payment outcomes across every rail, free for anyone to adopt.
Why this can't wait
Three clocks are already running.
Finance
Finance teams can't attribute machine spend.
Insurance
Insurers are writing AI out of liability coverage.
Regulation
The EU AI Act's record-keeping rules arrive in December 2027.
All three need the same thing: a trustworthy record of what autonomous software did with money.
That record is too important to sit inside one rail's database.
It should be verifiable by anyone.